Overhead-Aware Composition of Parallel Components Using Lazy-Parallel Function Calls

Parallel programming languages, APIs and tools available today try to extract as much performance as possible from parallel programs. But when components are parallelized this way, and are composed inside parallelized sections of code, multiple instances of the component are created, each creating more parallel tasks, resulting in much higher overhead. In this paper we propose a method which forces parallel components to execute serially if parallel execution would cause a high overhead of execution. Three variants of the technique are presented, and simulation results used to gain an insight into the efficacy of the different techniques are discussed.